Comment by irthomasthomas

Have you seen the news about decrypting the hidden COT in U.S. models? [0] The decoded logs revealed instances where Claude memorized answers to test questions beforehand while making its final output look like it had derived the answer step-by-step—hiding the memorization from the user.

0: https://www.alphaxiv.org/abs/2608.09867?hl=en-GB